"A microrobot that can truly operate inside a human body needs more than just a clever material or a smart algorithm," explains Dr. Zhang. "It needs to sense and adapt through its physical body, leverage living components for natural biocompatibility, use computation to learn and plan, and remain under expert human supervision where safety is critical. Each domain alone has severe limitations, but together they can produce robust, adaptive behavior."
Physical intelligence (PI) embeds sensing, actuation and adaptation directly into the robot's material and structure. Stimulus‑responsive hydrogels that release drugs at low pH, magnetoelastic composites that change shape in a magnetic field, and bioinspired helical tails that swim efficiently in viscous fluids all exemplify PI. "The material itself becomes part of the control loop," says Dr. Zhang. However, PI alone struggles with selectivity and multifunctionality in complex physiological environments.
Biological intelligence (BI) integrates living cells – stem cells, neutrophils, bacteria or algae – as active components. These provide built‑in chemotaxis, self‑propulsion, and even therapeutic functions. For example, algae‑based microrobots have been used to deliver drugs to lung metastases, and platelet‑coated systems exploit receptor‑mediated adhesion for tumour targeting. "BI offers unrivalled biocompatibility and environmental sensing, but comes with variability, short lifespan, and limited external controllability," notes Dr. Zhang.
Computational intelligence (CI) brings data‑driven perception, planning and learning. Deep learning estimates a microrobot's 3D pose from microscopy images, reinforcement learning (RL) learns to navigate complex channels, and digital twins accelerate algorithm training. Recent work has shown RL policies that position magnetic microrobots with sub‑30 μm error, or model‑based RL that guides ultrasound‑driven swimmers through vascular‑like microfluidic channels with over 90% success. "CI enables active adaptation, but it requires robust sensing and safe sim‑to‑real transfer," Dr. Zhang adds.
Human intelligence (HI) remains the ultimate supervisor. Through teleoperation, haptic shared control, augmented reality (AR) and virtual reality (VR) interfaces, clinicians can inject their expertise, enforce safety boundaries, and handle unexpected situations. One study showed that haptic shared control improved both speed and accuracy when steering multiple untethered microrobots in 3D. "Human oversight is non‑negotiable for medical microrobots – autonomy must be accountable," stresses Dr. Zhang.
The review systematically analyses how these domains can be coupled. PI–BI synergy appears in three forms: bioinspired structural designs (helical tails that improve propulsion), bioinspired functional materials (pH‑responsive hydrogels that release cargo on demand), and hybrid bio‑physical systems (living algae coated with magnetic nanoparticles for steerable drug delivery). PI–CI synergy uses field‑embodiment feedback loops – RL that directly updates magnetic coil currents based on tracked robot pose – and bidirectional learning where physics informs perception and machine learning optimizes embodiment.
BI–CI synergy tackles the challenge of tracking biohybrid agents in dense, low‑contrast tissue. A motion‑enhanced deep tracker (MEMTrack) achieves micron‑scale localization of bacteria in collagen, while RL has been used to rediscover chemotaxis‑like navigation policies. CI–HI synergy is realized through three interfaces: task/strategy (large language models translate natural‑language instructions into control policies), shared control (haptic devices blend human intent with autonomous obstacle avoidance), and perception (AR/VR overlays that fuse imaging data into an intuitive 3D scene).
The most advanced systems now combine three domains. For example, endoscopy‑assisted magnetic navigation of stem‑cell spheroids (PI–BI–CI) achieved rapid endoluminal delivery over 100 cm in less than eight minutes, with imaging verification. In knee‑joint cartilage regeneration, CI‑informed electromagnetic actuation boosted on‑target placement of living therapeutics from just 1.5% (injection only) to over 80%. MRI‑guided microrobot navigation in human‑scale hepatic arteries (PI–CI–HI) showed that computational posture planning and real‑time steering under clinician supervision increased targeting efficiency compared to standard injection workflows.
Despite these advances, major obstacles remain. Integrating smart materials, living cells and miniaturized electronics into a single, reproducible device is extremely challenging. Most prototypes are still handmade. Control must cope with nonlinear dynamics, biological variability, and noisy imaging – deep RL offers promise but demands extensive training, and simulations cannot yet capture all physiological complexity. Safety and ethics are paramount: autonomous microrobots performing medical acts will require clear regulatory pathways, fail‑safe mechanisms, and human oversight. Finally, the field lacks standardized benchmarks – comparing a helical swimmer to a biohybrid alga is not straightforward.
The authors outline four future directions: embodied design and manufacturing using computational optimization and multimaterial 3D printing to co‑evolve morphology, materials and control; intelligent swarms and onboard autonomy via distributed architectures and synthetic biology circuits; human‑robot shared control with AR/VR interfaces and constraint‑based assistance; and clinical translation through digital‑twin validation, standardized organ‑on‑a‑chip platforms, and regulatory frameworks adapted to combination products.
"Embodied cross‑domain intelligence is not just a checklist – it is about designing the interfaces and feedback loops that let different forms of intelligence work together seamlessly," concludes Dr. Zhang. "We believe this framework will guide the next generation of microrobots: devices that are not only smart but also safe, adaptable, and clinically trustworthy."
Authors of the paper include Zongcai Tan, Lan Wei, Kangyi Lu, and Dandan Zhang.
This work was supported by the Imperial-CNRS Collaboration Fund (Imperial College London–Centre National de la Recherche Scientifique, CNRS).
The paper "Embodied Cross-Domain Intelligence in Biomedical Microrobots: A Review" was published in the journal Cyborg and Bionic Systems on May. 14, 2026, at DOI: 10.34133/cbsystems.0546.
